Efficient Cross-Frequency Beam Prediction in 6G
Wireless Using Time Series Data
1st Vaibhav Bhosale§ 2nd Navrati Saxena
School of Computer Science Department of Computer Science
Georgia Institute of Technology San Jose State University
Atlanta, USA San Jose, USA
vbhosale6@gatech.edu navrati.saxena@sjsu.edu
4th Ada Gavrilovska
School of Computer Science
Georgia Institute of Technology
Atlanta, USA
ada@cc.gatech.edu
Abstract—Next generation 6G wireless envisions much higher
data rate and lower latency compared to 5G wireless networks.
Directional antennas, with narrow beams, across high mmWave
frequencies, hold the key to achieving such high data rates.
However, Beam Management (BM), which is the process of
finding appropriate transmit and receive beams, offers significant
challenges. Dynamic channel variation, user mobility, and narrow
beams in high frequency mmWave channels further complicate
these challenges. Efficient Machine Learning (ML) strategies can
be used to alleviate this overhead. In spite of their underlying
differences, sub-6 GHz and high frequency mmWave channels
share some similarities in array geometry, number of paths, and
surrounding environment. As sub-6 GHz channel characteristics
are relatively easier to acquire and learn, we introduce a new
machine learning framework, using transformer and LSTM to
learn sub-6 GHz channel information over time for efficient beam
prediction across high frequency mmWave channels. System
Level simulation results point out that when using time-series
data-based learning of beam patterns with transformers or
LSTMs in sub-6 GHz channels, our proposed scheme achieves
up to 99.5% top-5 beam prediction accuracy while reducing the
BM overhead by over 50% compared to existing work.
Index Terms—6G wireless, beam management, machine learn-
ing, transformer, LSTMs, channel measurements
I. I NTRODUCTION
The insatiable demand for high data rates and ubiquitous
connectivity is fueling the interests for upcoming sixth gener-
ation (6G) wireless. The vision of 6G wireless aims to provide
an order-of-magnitude more spectrum and ubiquitous connec-
tivity for emerging new applications, like eXtended Reality
(XR) and Metaverse, by exploring high-frequency millimeter
wave (mmWave) bands that provide the benefits of a big chunk
of available and un-explored frequencies. However, mmWave
bands face challenges such as high path loss, blockage,
and oxygen absorption. Dense network deployment with a
wide number of small-size, directional antennas is expected
to alleviate these challenges. This will be aided by beam
management (BM) to find the (near) optimal transmit/receive
§ This work was done as an intern at Mediatek Inc
3rd Ketan Bhardwaj
School of Computer Science
Georgia Institute of Technology
Atlanta, USA
ketanbj@gatech.edu
5th Abhishek Roy
Advanced Communication Technology
Mediatek Inc
San Jose, USA
Abhishek.Roy@mediatek.com
beams between the next-generation base station nodes (gNB)
and mobile User Equipment (UE).
One major challenge [1] in BM lie in inefficient beam
sweeping for the downlink (DL)/uplink (UL) beam corre-
spondence. Beam sweeping involves an exhaustive search to
find the beam/beam-pair having the strongest signal strength.
Naturally, performing an exhaustive search in all possible
directions is quite complex and expensive (both in terms of
latency and power). This already complex problem become
even more complicated in the mmWave bands, located in the
frequency spectrum ranging from 30 to 300 GHz. MmWave
bands provide the benefits of a big chunk of available and
un-explored frequencies. However, these benefits come with
a high attenuation cost in free space. To address the high
attenuation, a large number of highly directional MIMO
antennas are used, whose gain compensates for the path
loss. The use of these highly directional MIMO antennas
demands precise and efficient beam selection methods to
ensure the required application data rate and meet the strict
delay requirements. Another challenge such bands pose is
the low diffraction capacity and severe blocking caused by
most materials. Furthermore, beam searching across these
large number of narrow beams in high mmWave frequencies
incurs more delay and computational power. This makes BM
in mmWave more complex and expensive compared to BM in
sub-6 GHz bands. Therefore, 6G wireless will continue using
sub-6 GHz bands, also known as frequency range 1 (FR1)
for outdoor to indoor connectivity, while exploiting mmWave
bands or frequency range 2 (FR2) for high data rates in indoor
and outdoor environments [2].
Artificial intelligence (AI) and machine learning (ML) are
expected to play a key part in alleviating the complexity of
6G wireless. Major wireless standard bodies, like the 3rd
Generation Partnership Project (3GPP), anticipate native AI-
ML support in 6G wireless. 3GPP’s efforts encompass AI-
ML applications for networks and radio interfaces, with an
eventual goal of standardization [3]. More recently, 3GPP
has recognized the use of AI/ML techniques for efficient
BM as a key use-case. As many fundamental relationships
in wireless communications are often non-linear, deep neural
networks (DNNs) have recently been adopted in various fields
of wireless communications including BM [4]–[12].
Interestingly, while different types of DNN models [11],
[12] have been recently used for beam prediction across
frequencies, time-series data for cross-frequency BM is not
yet explored. As BM in wireless involves non-linear mappings
across time-varying wireless channels, we believe that explor-
ing time-series data for cross-frequency BM has sufficient
potential for improving its efficiency and accuracy. This is
where we believe that models, such as transformers [13]
and long short-term memory (LSTMs) [6] come into play§ .
Similar to recurrent neural networks (RNNs) [5], LSTM is
a deep learning model architecture capable of learning long-
term dependencies for sequence prediction problems. A trans-
former, on the other hand, is another deep learning model
architecture that avoids recurrence and instead relies entirely
on an ‘attention’ mechanism, differentially weighing the sig-
nificance of each part of the input data for obtaining global
dependencies between input and output. Removing recurrence
in favor of attention mechanisms endows transformers with
significantly higher parallelization, compared to other popular
DNN methods like RNNs and LSTMs.
In this paper, we introduce a new time-series data-based
beam prediction using models, such as LSTMs and trans-
formers that exploits channel characteristics of a sub-6 GHz
channel to choose a mmWave beam. The use of LSTMs
incurs lower number of FLOPs resulting in lower energy
consumption, while transformers can serve inference faster due
to the presence of its underlying attention mechanism. Our
framework lets 6G developers choose which model they need
to use based on their specific requirements. Assuming a sub-6
GHz connection is already established, we learn its underlying
channel information over time by using LSTM/transformer
learning. While learning over sub-6 GHz information aids
in reducing the search space needed for the initial mmWave
beam establishment, the use of LSTMs/transformers over time
series helps in the efficient capturing of time-varying wireless
channel characteristics.
The rest of the paper is organized as follows: In Sec-
tion II we take a look into the major related works on beam
prediction and point out our motivation behind using time
series information. Section III explains the system model and
problem formulation. Subsequently, we explain our proposed
time-series based beam prediction method in Section IV.
Simulation results in Section V show the efficacy of our
solution. Section VI concludes the paper with pointers to
future research.
II. R ELATED W ORKS AND M OTIVATION
BM problem has recently raised an increased research
interest, especially for the high frequencies (e.g., mmWave
and THz) 6G wireless networks. These high frequency net-
§ RNN was also a candidate, but since it experiences vanishing/exploding
gradients for longer sequences, it does not help here.
works are known to suffer from significant path loss and
blockage [14], and hence dense network deployments with
large-scale antenna arrays are used to achieve high beam-
forming gains to compensate this loss. However, the increased
scale of antenna arrays introduces high overhead when UEs try
to periodically measure the signal quality of different beams
for reporting the beam measurements for beam alignment and
potential handovers [15].
Adaptive learning of time varying wireless channels using
AI/ML techniques are recently used to address this challenge
and enable reliable and efficient beam management [16], [17].
The work in [4] uses reinforcement learning to develop a
multi-armed bandit framework for beam tracking. A data-
driven strategy that fuses the reference signal received power
(RSRP) with orientation information using an RNN is devel-
oped in [7]. The authors in [8] use LSTM to support large-
scale antenna array enabled hybrid, directional BM in hotspot-
based virtual small cells. [9] explores LSTM to utilize the
past channel state information (CSI) for efficient prediction of
future channels in vehicular mmWave systems. Deep learning
techniques, similar to Natural Language Processing (NLP)
is used in [10] to predict the best serving beams of the
mobile UEs. Interestingly, the work in [11] points out the
efficacy of fully-connected neural networks (FCNN) to closely
approximate these mapping functions, required for predicting
the optimal mmWave beams from sub-6 GHz channels. [12]
also uses deep neural networks for exploring power delay
profile (PDP) of a sub-6 GHz channel to predict the optimal
mmWave beams.
Broadly a lot of this work has focused on predicting beam
parameters by either (i) looking at fewer beam parameters, (ii)
parameters over time, (iii) parameters in a different frequency
band, etc. Our focus is on the latter two classes of problems,
where AI/ML has been recently used to share information
along the temporal and frequency domains individually. How-
ever, to the best of our knowledge, exploiting time series data
for beam prediction in different frequency bands is not yet
explored. As BM in high frequency mmWave bands are more
complex and expensive compared to BM in sub-6 GHz bands
and wireless channels are inherently time-varying, we posit
that there is an additional value in combining the information
across the two domains (time and frequency) to use temporal
information for improving the prediction accuracy in the
frequency domain beam prediction. Fortunately, LSTMs and
transformers provide an efficient tool to capture such time
series information [18] Although originally developed for
Natural Language Processing (NLP) [6], [13], over the next
few sections we will show how time-series based learning can
delve into time varying wireless channels information to assist
and improve cross-frequency BM in 6G wireless.
III. S YSTEM AND C HANNEL M ODELS
We consider a system where uplink communication happens
in FR1 and downlink in FR2 as shown in Figure 1. We assume
MF R1 antennas for uplink and MF R2 antennas for downlink.
We follow a similar system operation and channel models as
defined in [11]. We use the uplink channel vector as hF R1 [k] ∈

FR1 hFR1 FR1
Transceiver Transceiver
FR2 hFR2 FR2
Transceiver Transceiver UE
Base-station
Fig. 1: Depicting the system setup where the FR1 and FR2
antenna arrays are colocated on the base-station and UE.
CMF R1 ×1 at the kth subcarrier with k = 1, . . . , K (K is the
total number of subcarriers) and the signal received at the gNB
sub-6 GHz or FR1 array as:
yF R1 [k] = hF R1 [k]sp [k] + nF R1 [k], (1)
where sp [k] is the uplink pilot signal satisfying E∥sp [k]∥2 =
PF R1
K , with PF R1 as the uplink transmit power, and nF R1 [k] ∼
NC (0, σ 2 , I) is the noise at the gNB FR1 array. Similarly,
defining the downlink beamforming vector as f ∈ CMF R2 ×1
leads to the signal received by the UE as:
yF R2 [k̄] = hTF R2 [k̄]f sd + nF R2 [k̄] (2)
where hF R2 [k̄] ∈ CMF R2 ×1 represents the downlink channel
from the mobile UE to the gNB FR2 array at the k̄ th
subcarrier, k̄ = 1, 2, . . . , K̄. Owing to the hardware constraints
in the FR2 analog beamforming vectors, these vectors are
generally selected from quantized codebooks. Thus, it can
be assumed that the beamforming vector f takes one of the
candidate values collected in the codebook F , i.e., f ∈ F ,
with cardinality ∥F ∥= NCB . We adopt a geometric (physical)
model for the FR1 and FR2 channels [19]. Using this allows
us to model FR2 channel as
DXC −1 X L
2πk
hF R2 [k] = αl e−j K p(dTs − τl )a(θl , ϕl ), (3)
d=0 l=1
where L is the total number of channel paths, α, τ, θ, ϕ are
the path gains (including the path-loss), the delay, the azimuth
angle of arrival (AoA), and elevation AoA, respectively, and
a is the steering vector for the lth channel path. Ts represents
the sampling time while DC denotes the cyclic prefix length
(assuming that the maximum delay is less than DC Ts ). Using
this channel model and the system setup mentioned above,
the achievable downlink rate for an FR2 channel hF R2 and
beamforming vector f can be written as:
K̄
X the time series nature of this data. We observe comparable
R({hF R2 [k̄], f}) = log2 (1 + SN R∥hTF R2 f∥2 ) (4)
k̄=1
PF R2
where SNR is defined as SN R = Kσ 2 . The optimal beam-
F R2
forming vector is the one that maximizes the R({hF R2 [k̄], f})
which we denote by f ⋆ . Predicting f ⋆ requires estimating the
channel hF R2 or an online exhaustive beam training, which
incurs large training overhead, especially for high-frequency
FR2 or mmWave channels.
[11], [12] have recently pointed out that it is feasible to
use the FR1 channel information to predict the beamforming
vector in FR2. Moreover, they have also shown that there exists
a mapping from the FR1 channels to achievable rate in FR2
channel.
Φn : hF R1 → R(hF R2 , fn ), n = 1, 2, . . . , ∥F ∥, (5)
and that there exists a mapping function that can be leveraged
to obtain the optimal mapping between FR1 and FR2 channels.
Let’s define the optimal mapping function as ηt⋆ at timestep t
as
η ⋆ = argmaxn∈{n=1,2,...,∥F ∥} ΦnF R1 (hF R1 ) (6)
The works in [7]–[10] showed that temporal FR2 channel
information can be utilized to predict the optimal FR2 beam-
forming vector at next time-step, thereby giving
Ψt : {ht−m t−m+1
F R2 , hF R2 , . . . , ht−1 t ⋆
F R2 } → R(hF R2 , f ) (7)
Combining with Equation 7 with Equation 6, gives us
Ψt : {η ⋆ (ht−m ⋆ t−m+1
F R1 ), η (hF R1 ), . . . , η ⋆ (ht−1
F R1 )} (8)
t ⋆
→ R(hF R2 , f )
We term this composition Ω mapping from the FR1 chan-
nels for m time-steps to the optimal beamforming vector at
the latest time-step.
Ωt : {ht−m t−m+1
F R1 , hF R1 , . . . , ht−1 t
F R1 } → R(hF R2 , f )
⋆
(9)
Prior works have shown that high beam training time
raises significant challenges with solving these functions using
traditional (non-ML) approaches. Hence, we decided to look
at machine learning techniques that can help to solve this
problem. We observe that this is quite similar to the mapping
used to solve time-series problems [18] in AI/ML. Different
ML approaches, like RNNs, LSTMs, and transformers have
been used to solve these types of time-series problems. We
next look at which of these approaches will be suitable for
this problem.
IV. C ROSS F REQUENCY BM USING T IME S ERIES DATA
As we discussed earlier, ML can help to drastically reduce
the time associated in BM by incorporating information about
the channel and predicting the best k beams. Prior work [11],
[12] focused on using the FR1 measurements at the current
timestep to exploit the correlation between the FR1 and
FR2 measurements. We observed that this correlation can
be composited with the correlation between measurements
at consecutive time steps i.e., a time series data of FR1
measurements helps with better prediction of FR2 beams.
The most common techniques to deal with time series
data include RNNs, LSTMs and transformers. We ruled out
RNNs as a potential solution since they suffer from vanish-
ing/exploding gradients [6] and hence won’t be able to capture
accuracy results for transformers and LSTMs (Section V) with
the transformer doing marginally better. However, transformers
have a faster inference speed due to their ability to replace
recurrence to entirely focus on attention making them faster
while also increasing the parallelism of the model. On the
other hand, LSTMs incur less number of Floating Point
Operations (FLOPs) [20], [21] that will result in lower energy
utilization to run the model. Our contribution in this paper is
 Layer Output Shape Parameters Dropout Layer
Layer Output Shape Parameters
Input (10,504) 0
Input (10,504) 0
Transformer Block 1 (10,504) 256159 Normalization Layer
LSTM 1 (10,256) 779264
Transformer Block 2 (10,504) 256159
LSTM 2 (10,256) 525312 Feed-Forward Layer
Transformer Block 3 (10,504) 256159
LSTM 3 (10,128) 197120
Transformer Block 4 (10,504) 256159
LSTM 4 (10,64) 49408 Feed-Forward Layer
Average Pooling (504,) 0
LSTM 5 (10,32) 12416
Dense (128,) 64640
Dense (128,) 4224 Multi-Head Attention
Dropout (128,) 0
Dense (504,) 65016
Dense (504,) 65016 Input

Fig. 2: The structure of every

TABLE I: The different layers used in TABLE II: The different layers used individual transformer block
the design of the LSTM model in the design of transformer

to provide a framework to use time-series data, so we provide

the various options through which this data can be utilized
for beam management, while also outlining the tradeoffs
associated with the same.
• Model Structure: Table I and Table II show the structure
of our LSTM and transformer models. We use standard
LSTM blocks of the Keras library. The LSTM model
consists of 5 LSTM cells and the transformer model
consists of 4 transformer blocks (Figure 2) to process
sequential data. These layers are followed by dense layers
that help transform the obtain the FR2 data values for the
requisite channels (504 in our current implementation)
that help us identify the k best beams for BM.
• Training Procedure. Both these models are trained using
the data generated by our system-level simulation (more
details in Section V). We use sparse categorical cross- Fig. 3: The layout of different cells used in our simulation
entropy as the loss function and the Adam optimizer with environment
a learning rate of 0.0001 for training. The training process
for both the models required 40-60 epochs for our dataset.
The design of our algorithm provides flexibility to choose
Algorithm 1 Beam Management function for cross-frequency the value of k which provides a tradeoff between the mea-
prediction surement overhead and accuracy. As we show in Section V,
the top-5 beams can provide over 99% accuracy using LSTMs
1: function C ROSS F REQUENCY BM(FR1Data) and transformers.
2: P rocessedData ← DataP rocessing(F R1Data) V. S IMULATION E XPERIMENTS AND R ESULTS
3: topKBeams ← T imeSeriesM odel(P rocessedData)
4: topKM easurements ← M easure(topKBeams) In this section, we first briefly present our simulation
5: return ArgM ax(topKM easurements) setup and go through the overall simulation methodology.
6: end function Subsequently, we discuss the simulation results obtained by
carrying out the simulation experiments.
We describe the steps involved in our cross-frequency beam A. Simulation Setup.
management algorithm in Algorithm 1. At every step, the FR1 As ML-enabled 6G modem chipsets are not yet widely
data for the previous 10 timesteps is collated to act as the input available for experimental design and verification, we resort
to the algorithm. to MediaTek’s custom-built System Level Simulator (SLS)
1) First, the FR1 data is normalized which helps to maintain to generate data using 3D ray tracing and subsequently use
the numerical stability of the data. 3GPP-specified simulation parameters [22]. We use the two
2) Next, we run this data through a trained time-series frequencies 4 GHz (for sub-6GHz/FR1) and 30 GHz (for
model (transformer / LSTM) to provide the top-k beams. mmWave/FR2). Our SLS simulates an outdoor environment
3) These k beams are then measured, ensuring that a tiny consisting of 21 cells (as shown in Figure 3) for 1000
subset of all the beams needs to be measured thus, timesteps, with each timestep 5 ms apart. At every cell, we
speeding up the beam management process. consider a base station equipped with two co-located uniform
4) Finally, the optimal beam from these k beams is chosen. linear arrays for both frequencies. We run our simulations for
 Top 1 Top 3 Top 5

FLOPs (millions)
100 99.1 97.96 99.34 99.54 200
95.42 96.5 96.75
92.76 150
90 100
Accuracy

80
50
0
72.42 FCNN TF#1 TF#5 TF#10 TF#20 TF#40
70 66.52
69.69
Model
63.3
60 Fig. 5: Comparing the number of Floating Point Operations
FCNN RNN LSTM Transformer
Model required for all the different models.

Fig. 4: Top-1, Top-3, and Top-5 accuracy for different models.

Top 1 Top 3 Top 5
100 99.22 99.41 99.44 99.54 99.56
96.26 96.79 96.69 96.75 97.07
a total of 100 UEs. We consider both stationary and mobile
UEs (upto 30km/h). 90

Accuracy
B. Simulation Methodology.
80
We compare our methodology of using time-series data with 71.91 72.42 71.44
69.97 71.28
prior work [11], [12] that suggested the use of fully connected 70
neural networks (FCNN) to determine the best beam in FR2.
The FCNN model consists of an input layer and four hidden 60
layers all having the same number of neurons. All the layers TF 1 TF 5 TF 10 TF 20 TF 40
are activated with the ReLu function [23] followed by the Model
dropout function [24], except the last layer which is activated
Fig. 6: Top-1, Top-3, and Top-5 accuracy for transformer
with the softmax function [25]. In addition to using the FCNN
variants. (TF n refers to TF with window size n)
model, we also show the results for a RNN model as a
baseline comparing with our proposed LSTM and transformer
models. All the models used in our simulation are developed Top 1 Top 3 Top 5
using Keras [26]. While we use the entire models for our 100 98.17 98.32 99.34 99.23 98.88
simulation study, using Keras provides us the benefit of using 96.5 95.58 94.89
93.91 94.14
the TensorFlow Lite Converter which can easily convert our
90
models to run efficiently on UEs.
Accuracy

We use a 90-10 train-test split for our simulations. During

the training phase, we use a 80-20 train-validation split. We
80
compare the models based on the Top-1 accuracy, Top-3 69.69
accuracy, Top-5 accuracy, and the number of floating point 70 65.81
68.48 67.06
65.6
operations (FLOPs) required for every instance of inference.
We obtain the FLOPs for the FCNN and transformer models 60
LSTM 1 LSTM 5 LSTM 10 LSTM 20 LSTM 40
using the keras-flops library [27]. However, as there is no way Model
to find the FLOPs for RNN and LSTM, we do not share the
FLOPs for those models. Fig. 7: Top-1, Top-3, and Top-5 accuracy for LSTM variants.
(LSTM n refers to LSTM with window size n)
C. Results
We measure the accuracy for the scenario using all 21 cells
from our simulation environment in Figure 4. We observe transformer is due to the inherent design of the two models.
that the transformer model outperforms all the other models In the case of an FCNN model, the number of multiply-
(FCNN, RNN and LSTM). While RNN performs poorer com- accumulate operations increases as the number of FLOPs
pared to the FCNN model due to its difficulties in capturing increases quadratically with the increase in the number of
patterns in longer sequences, LSTM does indeed perform neurons [28] whereas, for a transformer, the number of FLOPs
better than the FCNN, but slightly worse than the transformer. increases linearly. Thus, this has greater implications while
Further, as we see in Figure 5, a transformer requires far fewer modifying the size of the model, which is why we advocate
number of FLOPs compared to the FCNN. However, as shown the use of a transformer.
in [20], [21], LSTMs will require even fewer FLOPs. Further, we also show a micro-benchmark for the trans-
This difference in trends for the FLOPs for the FCNN and former varying the window sizes trying to converge to an
optimal window size. We observe that the accuracy of the
model increases rapidly as we increase the window size from 1
to 10, with very little increase from 10 to 20, but then observes
a drop at window size 40, as shown in Figure 6. This tradeoff
is important especially since the number of FLOPs required
by the model increases as the window size increases as we
observed in Figure 5.
Similarly, we also show a micro-benchmark for LSTMs
while varying the window size. Similar to the transformer, the
accuracy of the model increases while increasing the window
size from 1 to 10, but then drops at 20 and 40 (Figure 7). In
a way, LSTMs are constrained to utilize less amount of data
compared to the trnasformer.
VI. C ONCLUSION
In this paper, we introduce a new ML framework which
explores using time series based transformers and LSTMs to
efficiently learn the historical channel measurement of sub-6
GHz wireless channels. As mmWave channel characteristics
are more complex and expensive to learn, we subsequently
use this sub-6 GHz wireless channel characterisitics, learnt
over time, to efficiently predict the channel information and
BM in high frequency mmWave channels for 6G wireless.
System Level simulation results demonstrate that our time
series based learning of beam patterns in sub-6 GHz channels
results in up to 99.5% top-5 accuracy for beam prediction and
reduces the BM overhead by over 50% in mmWave channels
compared to existing approaches. We hope that our work
inspires the community to incorporate such learning based on
time series data for other aspects of beam management and
channel estimations.
R EFERENCES
[1] Y. Heng, J. G. Andrews, J. Mo, V. Va, A. Ali, B. L. Ng, and J. C. Zhang,
“Six Key Challenges for Beam Management in 5.5G and 6G Systems,”
IEEE Communications Magazine, vol. 59, no. 7, pp. 74–79, 2021.
[2] Q. Xue, C. Ji, S. Ma, J. Guo, Y. Xu, Q. Chen, and W. Zhang, “A Survey
of Beam Management for mmWave and THz Communications Towards
6G,” arXiv preprint arXiv:2308.02135, 2023.
[3] 3GPP Release 18, “Summary for RAN rel-18 package, 3GPP RAN
Plenary 94-e,” Technical Report RAN Chair, RP-213469, 2022.
[4] I. Aykin, B. Akgun, M. Feng, and M. Krunz, “MAMBA: A Multi-armed
Bandit Framework for Beam Tracking in Millimeter-wave Systems,” in
IEEE INFOCOM 2020-IEEE Conference on Computer Communications.
IEEE, 2020, pp. 1469–1478.
[5] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning Internal
Representations by Error Propagation,” California Univ San Diego La
Jolla Inst for Cognitive Science, Tech. Rep., 1985.
[6] S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,” Neural
computation, vol. 9, no. 8, pp. 1735–1780, 1997.
[7] K. N. Nguyen, A. Ali, J. Mo, B. L. Ng, V. Va, and J. C. Zhang,
“Beam Management with Orientation and RSRP using Deep Learning
for Beyond 5G Systems,” arXiv preprint arXiv:2202.02247, 2022.
[8] Y. Liu, X. Wang, G. Boudreau, A. B. Sediq, and H. Abou-zeid, “Deep
Learning Based Hotspot Prediction and Beam Management for Adaptive
Virtual Small Cell in 5G Networks,” IEEE Transactions on Emerging
Topics in Computational Intelligence, vol. 4, no. 1, pp. 83–94, 2020.
[9] Y. Guo, Z. Wang, M. Li, and Q. Liu, “Machine Learning Based mmWave
Channel Tracking in Vehicular Scenario,” in 2019 IEEE International
conference on communications workshops (ICC Workshops). IEEE,
2019, pp. 1–6.
[10] A. Ö. Kaya and H. Viswanathan, “Deep Learning-based Predictive
Beam Management for 5G mmWave Systems,” in 2021 IEEE Wireless
Communications and Networking Conference (WCNC). IEEE, 2021,
pp. 1–7.
[11] M. Alrabeiah and A. Alkhateeb, “Deep Learning for mmWave Beam and
Blockage Prediction Using Sub-6GHz Channels,” IEEE Transactions on
Communications, vol. 68, no. 9, pp. 5504–5518, 2020.
[12] M. S. Sim, Y.-G. Lim, S. H. Park, L. Dai, and C.-B. Chae, “Deep
Learning-Based mmWave Beam Selection for 5G NR/6G With Sub-6
GHz Channel Information: Algorithms and Prototype Validation,” IEEE
Access, vol. 8, pp. 51 634–51 646, 2020.
[13] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez,
L. Kaiser, and I. Polosukhin, “Attention is all you need,” Proceedings
of the 31st International Conference on Neural Information Processing
Systems (NIPS’17, pp. 6000–6010, 2017.
[14] Y. Niu, Y. Li, D. Jin, L. Su, and A. V. Vasilakos, “A Survey of
Millimeter Wave (mmWave) Communications for 5G: Opportunities and
Challenges,” Wireless networks, vol. 21, no. 8, pp. 2657–2676, 2015.
[15] S. Hur, T. Kim, D. J. Love, J. V. Krogmeier, T. A. Thomas, and
A. Ghosh, “Millimeter Wave Beamforming for Wireless Backhaul and
Access in Small Cell Networks,” IEEE transactions on communications,
vol. 61, no. 10, pp. 4391–4403, 2013.
[16] M. Q. Khan, A. Gaber, P. Schulz, and G. Fettweis, “Machine Learning
for Millimeter Wave and Terahertz Beam Management: A Survey and
Open Challenges,” IEEE Access, vol. 11, pp. 11 880–11 902, 2023.
[17] K. Ma, Z. Wang, W. Tian, S. Chen, and L. Hanzo, “Deep Learning
for Beam-Management: State-of-the-Art, Opportunities and Challenges,”
arXiv preprint arXiv:2111.11177, 2021.
[18] S. Ahmed, I. E. Nielsen, A. Tripathi, S. Siddiqui, G. Rasool, and R. P.
Ramachandran, “Transformers in Time-series Analysis: A Tutorial,”
arXiv preprint arXiv:2205.01138, 2022.
[19] R. W. Heath, N. Gonzalez-Prelcic, S. Rangan, W. Roh, and A. M.
Sayeed, “An Overview of Signal Processing Techniques for Millimeter
Wave MIMO Systems,” IEEE journal of selected topics in signal
processing, vol. 10, no. 3, pp. 436–453, 2016.
[20] W. Li, J. Qin, C.-C. Chiu, R. Pang, and Y. He, “Parallel Rescoring
with Transformer for Streaming On-Device Speech Recognition,” arXiv
preprint arXiv:2008.13093, 2020.
[21] Y. Wang, Q. Wang, S. Shi, X. He, Z. Tang, K. Zhao, and X. Chu,
“Benchmarking the Performance and Energy Efficiency of AI Accelera-
tors for AI Training,” in 2020 20th IEEE/ACM International Symposium
on Cluster, Cloud and Internet Computing (CCGRID). IEEE, 2020, pp.
744–751.
[22] 3GPP Release 14, “Study on International Mobile Telecommunications
(IMT) parameters for 6.425 - 7.025 GHz, 7.025 - 7.125 GHz and 10.0
- 10.5 GHz,” 3GPP Technical Report, 2018.
[23] A. F. Agarap, “Deep Learning using Rectified Linear Units (ReLU),”
arXiv preprint arXiv:1803.08375, 2018.
[24] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and
R. Salakhutdinov, “Dropout: A Simple Way to Prevent Neural
Networks from Overfitting,” Journal of Machine Learning Research,
vol. 15, no. 56, pp. 1929–1958, 2014. [Online]. Available:
http://jmlr.org/papers/v15/srivastava14a.html
[25] J. S. Bridle, “Probabilistic Interpretation of Feedforward Classification
Network Outputs, with Relationships to Statistical Pattern Recognition,”
in Neurocomputing. Springer, 1990, pp. 227–236.
[26] F. Chollet et al. (2015) Keras. [Online]. Available:
https://github.com/fchollet/keras
[27] Keras Flops. [Online]. Available: https://pypi.org/project/keras-flops/
[28] M. Hollemans, “How fast is my model?”
https://machinethink.net/blog/how-fast-is-my-model/.